ays-v0 implements the AYS model first described in Kittel et al. that phenomenologically models emissions, economic output and growth of renewable energy. This environment follows the implementation that Strnad et al. used to put the AYS model into a OpenAI gym environment.
Observation Space The agent observes the variables A, Y and S. A is the excess atmospheric carbon stock. Y is economic output. S is the renewable energy knowledge stock variable.
Model Dynamics The model is a system of ODE’s which are described on p. 15 of Kittel et al.
Action Space The agent has 4 available actions: do nothing, lower economic growth, levy a fossil fuel tax or both lower economic growth and levy a fossil fuel tax.
Reward Function The agent is given a constant reward if it keeps the system within some planetary boundaries.
dice-v0 implements Nordhaus’ DICE model which is a richly detailed integrated assessment model. This implementation closely follows the pyDICE model, which is a python version of the originally published DICE model.
Observation Space The agent observes 29 continuous valued variables that describe various economic and geophysical properties.
Model Dynamics Dynamics follow the DICE model – see the user guide.
Action Space The agent controls 2 continuous valued variables: \(\mu\), the emissions reduction rate, and \(S\), which is the savings rate.
Reward Function The reward function is intricate. Details are given in the user guide, but generally, the agent is rewarded for improving social welfare.